Employee supervises its agents

The idea

The AI employee plans, briefs, reviews, integrates and talks to people. It hands bounded plan slices to cheaper models or agent CLIs.

The problem
A top-tier model spends expensive turns on work that a cheaper model can do.
Why it matters
Strong reasoning stays on the main line, and slices run on the cheapest model that fits.
Where it stands
In production as of

A step-by-step replay of this idea is not published yet. The library adds one when a recorded run shows it.

  • Native implementation round

    Implementation runs as one sandboxed Claude Code or Codex session, briefed by a compiled file. Its tools go through the gateway hook. The supervisor gets back only the report, the diff and the checks.

  • Context contract per role

    Each model call carries only its role's layers, in a fixed order. People's messages enter as task state with their ids. Budgets are kept by folding and ranking, never by cutting the prefix.

  • Internal knowledge first, then specialists

    A worker needs internal evidence before its first web search or question: a knowledge or memory read, a colleague's answer, or a recorded miss. After that it can consult read-only specialist sub-runs, search the web and, with approval, ask an outside expert.

  • Parallel work in one repository by footprint

    Each plan predicts the files and functions it touches; overlapping work goes to the same worker or waits, and disjoint work runs side by side without worktrees.

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.